As speech translation (ST) systems become increasingly prevalent, understanding their vulnerabilities is crucial for ensuring robust and reliable communication. However, limited work has explored this issue in depth. This paper explores methods of compromising these systems through imperceptible audio manipulations. Specifically, we present two approaches: (1) adapting perturbation-based techniques used for automatic speech recognition (ASR) attacks to the ST context, making our work the first to apply this approach to ST, and (2) proposing a novel music generation-based method to guide targeted translation, while also conducting more practical over-the-air attacks in the physical world. Our experiments reveal that carefully crafted audio perturbations can mislead translation models to produce targeted, harmful outputs, while adversarial music achieve this goal more covertly, exploiting the natural imperceptibility of music. These attacks have proven effective across multiple languages and translation models, highlighting a systemic vulnerability in current ST architectures. Beyond immediate security concerns, our findings highlight broader challenges in the robustness and interpretability of neural speech systems.

Speaker verification (SV) constitutes a prevalent biometric recognition technology safeguarding the security of sensitive data and regulating access to vital infrastructure. Current audio adversarial attack researches predominantly concentrate on deceiving systems through an touching on inference parameters of the underlying model architectures. This work introduces a pragmatic and formidable attack scenario wherein adversarial perturbations are introduced during the SV enrollment phase. These meticulously engineered distortions subtly alter voice embedding extraction, consequently influencing the verification outcome for the adversary. To enable this assault, we posit perturbations comprised of inaudible high-frequency noise and optimize them within an energy-based framework. Under black-box conditions, an evolutionary algorithm, differential evolution, resolves the intricate optimization challenge. Furthermore, we leverage a spectrum of realistic constraints, including absolute hearing thresholds, signal-to-noise ratios, and over-the-air propagation effects, to augment the robustness of our methodology. Rigorous experimentation across four voice identity datasets affirms the efficacy of our approach, demonstrating success rates exceeding 90% in digital environments and 78% in physical settings against four widely employed SV models.